AI for Drug Discovery: A New Era with Isomorphic Labs

For over a decade, artificial intelligence has been touted as a catalyst to dramatically accelerate the discovery of new drugs. Despite billions of dollars in investment, the number of AI-designed medicines that have reached patients remains limited. This is partly due to the inherently long timelines for clinical testing and the intrinsic complexity of drug development. In this scenario, Isomorphic Labs, a Google DeepMind spinout, is emerging as a key player, capitalizing on DeepMind's Nobel Prize-winning work on protein structure prediction.

The company has already signed major drug-discovery partnerships with industry giants like Novartis and Eli Lilly, and recently raised $2.1 billion in funding. In February, it published a technical report describing its new Isomorphic Drug Design Engine (IsoDDE), a system designed to identify "pockets" on proteins where drugs can bind and, more generally, to predict how proteins and drug molecules interact. This approach promises to overcome some of the limitations that have so far hindered the widespread adoption of AI in the sector.

Beyond AlphaFold: IsoDDE's Precision in Molecular Modeling

DeepMind's work with AlphaFold2 and AlphaFold3 represented a massive leap forward for computational biology. AlphaFold2, in particular, was recognized with the Nobel Prize for arguably solving the problem of protein folding. However, as Adrian Stecuła, a group leader in the machine learning organization at Isomorphic Labs, points out, proteins do not exist in a vacuum; they interact with a wide variety of other cellular biomolecules, including nucleic acids, small molecule ligands, ions, and other proteins. AlphaFold3 extended this capability, allowing all these interactions to be modeled within a single framework.

Despite these advancements, evaluations following AlphaFold3's release highlighted a crucial limitation for drug discovery: the model's performance decreases when dealing with novel protein pockets or those significantly distant from the training set. For drug discovery, it is essential to explore novel mechanisms of action, which often involve targeting never-before-observed pockets. IsoDDE addresses this challenge, not only predicting where a ligand binds to a protein, but also how it binds, how tightly it binds, and a plethora of other properties crucial for developing an effective drug. IsoDDE's unified system supports structure prediction, pocket identification, and binding affinity prediction.

Identifying Cryptic Pockets and Implications for AI Infrastructure

A key example of IsoDDE's capability is the identification of a "cryptic pocket" on the cereblon protein, a fundamental component in the targeted protein degradation pathway. Cryptic pockets are cavities on protein surfaces that are not immediately obvious in the unbound state of the protein but only open upon the binding of the right ligand. In a recent study published in Nature, a completely novel cryptic pocket on cereblon was described. IsoDDE was able to perfectly predict the location of this pocket using only the protein sequence as input, and accurately recapitulate the crystal structure of ligand binding, for both orthosteric and allosteric ligands.

This ability to identify and model interactions at unconventional binding sites is crucial. Many diseases are associated with proteins that, despite being known targets, do not present pockets that can be easily "drugged" by traditional medicines. IsoDDE expands the therapeutic toolkit, not limited to small molecules, but also extending to antibodies, molecular glues, and peptides. For companies operating in this sector, implementing such sophisticated computational systems entails significant infrastructure requirements. Managing complex models and processing vast biological datasets demand high computational capabilities, often relying on GPUs with large amounts of VRAM and robust storage and networking infrastructure. For those evaluating on-premise deployments, considering TCO and data sovereignty is essential, as these are critical aspects for compliance in the pharmaceutical industry. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these trade-offs.

The Automated Future of Drug Discovery

Despite the excitement surrounding AI in drug discovery, it is crucial to avoid the misconception that the sole ability to accurately model protein structure equates to solving the entire problem. As Stecuła emphasizes, a unified system like IsoDDE, with a plethora of different endpoints, is necessary to effectively model these complex systems. Isomorphic Labs is committed to continuously improving performance on the disclosed endpoints and developing new ones.

The long-term vision is an increasingly automated drug discovery process, where AI systems generate hypotheses, test them, and analyze results. Max Jaderberg, president of Isomorphic Labs, nicely framed this future by discussing "agentic workflows" in drug discovery. This prospect of advanced automation, though still evolving, suggests a paradigm shift in how new therapies will be identified and developed, with profound implications for AI infrastructure investment strategies.